RooflineAI's EIC Accelerator Compiler for Every Edge Chip

RooflineAI's EIC project starts from the claim that the tools for running trained AI models on phone chips, car controllers and factory sensors have not kept up. The CORDIS summary of RooflineAI's EIC project puts the problem in two numbers: more than 400,000 new AI models a year and hundreds of chip architectures, with existing tools that struggle to support new model types such as large language models and fail to use the hardware fully. RooflineAI, a 2024 spin-off of RWTH Aachen University, sells a compiler that takes a model from PyTorch, TensorFlow or ONNX and produces code for whatever processors sit on a chip.

The EIC Accelerator selected the project, ROOFLINE, at its March 2025 cut-off for blended finance. HiPEAC reported a €2.5 million grant plus a pre-committed equity investment for the company's next round, with 40 companies chosen from nearly 1,000 applicants. CORDIS records an EU contribution of €2,494,448 for a project running from 1 January 2026 to 31 December 2027. The project is ten months old, so this article records what was promised and what can be checked so far.

Compiling a model instead of interpreting it

The established route to edge deployment, which the company describes as a legacy stack based mainly on TensorFlow Lite, interprets a model at run time and partly depends on hand-written, manually optimised kernels for each chip. The company says this limits support for newer models such as language models. Its SDK instead compiles the model, generating code for each processor on a system-on-chip, and the company says deployment takes one line of Python.

The compiler is not built from scratch. The company website says it is based on MLIR and IREE, the open-source compiler and runtime that has been developed since 2019 and joined the Linux Foundation's LF AI and Data Foundation in May 2024. Supported targets listed on the site include Arm Cortex CPUs and Mali GPUs, Qualcomm, Apple and Google Tensor processors, x86, RISC-V, Nvidia RTX and NXP's Neutron NPU. Electronic Specifier carried the same description in August 2024 of the SDK as a single backend for CPUs, MPUs, MCUs, GPUs and AI accelerators.

Building on open source is both the strength and the commercial question. In August 2026 CTO Maximilian Bartel became chair of IREE's technical steering committee, and the company wrote in October 2026 that LFX Insights records more than 1,000 Roofline contributions to the project in the past year. The same post stresses that technical oversight stays with the committee as a whole. What Roofline sells is the layer on top: hardware enablement, integration, optimisation and support.

How the performance figures hold up

The grant summary on CORDIS says the SDK delivers up to 15 times lower latency and 35 times lower memory use than the industry standard, without naming that standard, the models or the hardware. No benchmark behind these two figures could be found on the company website or in the papers searched for this article.

The only figures set out in a paper are smaller. A May 2026 arXiv paper by Ege Beysel, Maximilian Bartel and Jan Moritz Joseph, all of Roofline, reports Arm SVE code that is competitive with, and often faster than, IREE's existing NEON output, by up to 1.45 times; the company's news page rounds this to 1.5 times and adds comparisons of up to 1.7 times against ExecuTorch and 6 times against TorchInductor. For NXP's chip, the company reports up to 3.2 times faster LLM prefill on CPU plus NPU than on the CPU alone. None of these has been reproduced by an outside group in any source found.

Partners, but no named customer

The closest thing to a customer in public is NXP. On 1 March 2026 Roofline announced that "NXP and Roofline have teamed up" to run large language models on the eIQ Neutron NPU of NXP's i.MX 95 processor, including models larger than the NPU's 2 GB of memory. The post contains no statement from NXP and does not describe a contract, and web searches for this article found no NXP statement that mentions Roofline. In June 2026, per the same news page, CEO Jan Moritz Joseph sat on a panel on European AI sovereignty at Edge AI London with NXP, Infineon, STMicroelectronics and Innatera.

Other links are of the same kind. Arm lists Roofline in its partner ecosystem catalogue as a development tools and AI software partner, which is a listing, not a contract. Siemens' Cre8Ventures wrote in October 2024 that Roofline would become a partner in its automotive digital twin marketplace. The company also announced day-zero support for a Liquid AI model in July 2026 and a publicly funded project, AI2EDGE, with MachineWare and Fraunhofer IPT, backed by the EU and North Rhine-Westphalia's economy ministry, with no amount published. The website refers to "real customer use cases on production SoCs" without naming any.

The money, as far as it is public

No funding round has been announced by the company. Data providers disagree on what was raised. Caplight lists a seed round dated 25 June 2024 with AQAL, Arcas Ventures, First Momentum Ventures, Pawao Capital and Combination VC, and puts total funding at $5.73 million; it also lists a grant dated 4 July 2025, without saying whether the total includes it. CB Insights calls the latest round an angel round of May 2024 and lists Pawao, Onsight Ventures, AQAL, Combination VC and First Momentum. Serena lists Roofline in its own portfolio without an amount or date.

The register shows the shape of the cap table, not its price. North Data records share capital of €25,000 at registration, rising to €35,674 in July 2025, €35,741 in January 2026 and €36,482 in February 2026, each time with a new shareholder agreement, and 21 active shareholders. Whether any of these increases is the EIC Fund's pre-committed equity is not public: no EIC Fund investment in Roofline has been announced in any source found. North Data also records German federal funding of €248,728 from April 2025 for a sub-project on software support for edge AI systems with RISC-V; that grant could not be confirmed in a second source.

The EIC project: promised and checkable

1. An SDK supporting PyTorch and TensorFlow modelsExistsProduct described on the company website and by Electronic Specifier; it predates the grant.
2. Targets across CPUs, GPUs and acceleratorsIn progressCompany demos on Qualcomm, Arm, Apple and NXP silicon; NPU support beyond NXP is described as work with unnamed vendors.
3. Up to 15x lower latency, 35x lower memory useClaimNo benchmark published; the only paper reports up to 1.45x on Arm SVE.
4. Commercial uptake in autonomous systems, industry and healthcareNot shownNo customer named; CORDIS shows no periodic report yet.

Why the EU funds compiler software

The European Chips Act, in force since 21 September 2023, lists the design, manufacturing and packaging of advanced chips among its objectives. The CORDIS summary frames ROOFLINE's goal as European leadership in edge AI deployment, and the company's public work so far includes software support for NXP's chips and a panel on European AI sovereignty with NXP, Infineon and STMicroelectronics. Whether the market will pay a small company for a layer on top of an open-source project is the question the grant period has not yet answered.

RooflineAI was not asked for comment for this article. The reporting is based on publicly available sources, listed below.